• The spectral convolution of a graph is defined as multiplication of a node-wise signal with a convolutional filter where is the parameter of the filter in the Fourier domain. Thus

    Where represents the matrix of eigenvectors of the normalized Laplacian such that

  • can be understood as a function of the eigenvalues of the Laplacian.

  • One approximation to is to use Chebyshev Polynomials up to -th order.

    Where

    And we define the convolution as

    Where